# Anthropic says agents should dream offline

_Anthropic says offline dreaming lifted Claude from 13% to 48%, and small, tested memories beat large logs._

**Published:** 2026-11-08
**Source:** https://www.startuphub.ai/ai-news/ai-research/2026/anthropic-says-agents-should-dream-offline

---

[Anthropic Engineering](https://www.anthropic.com/research/components-of-agent-memory) says offline dreaming lifted Claude from 13% to 48% on 25 questions.

Agents now work across sessions. Logs pile up fast. Memory must persist without polluting context. [Anthropic](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/anthropic-restricts-ai-access-for-foreign-nationals) defines memory as experience kept between sessions and reread as context, per its post. That definition excludes weight changes, in-session learning and compaction.

Memory is a prediction. A saved rule bets a future session will need it. That bet paid off on Continual Learning Bench 1.0, per [Anthropic](https://www.startuphub.ai/ai-news/funding-round/2026/anthropic-s-965b-valuation-outshines-openai). Claude started each database question fresh but could read a persistent memory folder.

Newer models closed the learning loop. Claude Sonnet 4.6 left untested guesses scattered across files, per Anthropic. Claude Opus 4.7 consolidated notes and marked checked facts. Claude Opus 5 wrote rules, cited how each rule was verified and flagged traps. Claude Opus 5.5 and Sonnet 5.5 also closed the loop, with slightly higher baselines without memory.

The Pokémon Red test was starker. Claude Sonnet 3.7 wrote 413 files in 225 folders and won one badge, per Anthropic. [Claude Fable 5](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/anthropic-restricts-ai-access-for-foreign-nationals) wrote nine files in two folders plus global.md and won four badges. Sonnet 3.7 logged battles and map positions that went stale. Fable 5 kept goals, gym plans and tested rules. After seven failed balls on a Magikarp, it saved a timing trick. Opus 5.5 kept a similarly small folder and finished the game.

Strict instructions hurt stronger models. Telling Opus 5 and Opus 4.7 to record only failures cut much of the memory gain, per Anthropic. Older models needed steering. Prompting Sonnet 4.6 to read memory and test theories nearly doubled its CL-Bench score. Sonnet 5 already read memory in 9 of 10 sessions without prompting, compared with less than half for Sonnet 4.6.

Dreaming fixed errors that raw notes preserved. Near Mt. Moon, Claude mislabeled a PokéCenter and declared the cave done, per Anthropic. It then hit a one-way ledge three times and retreated to Pewter City. Replaying the transcript against the note corrected the location. With the fix, Claude re-entered the cave and reached Cerulean City. No memory and raw memory both failed the route.

The same offline pass built navigation. Over 128 dreaming sessions, Opus 5 agents explored about 350,000 Stainless files totaling roughly 420M tokens and built a linked wiki, per Anthropic. Practitioners already distilled that workflow, with a DEV Community explainer framing it as notes, salience judgments and dreaming. The cost is real: gains required repeated passes, and stale notes still mislead until checked.

---

Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.

© StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.